Search infrastructure is a corner where Rust is the implementation language rather than a binding: the vector databases, the full-text engine underneath hybrid search and the local embedding servers on these shelves are Rust programs, and they publish Rust clients alongside the Python ones most of their users reach for.
Assembling them into a pipeline is the part you still write yourself.
Vector databases and the ANN index crates underneath them. Several of the best-known vector databases are themselves written in Rust and publish first-party Rust clients.
A search engine you can point at a dataset and get relevance from in an afternoon — typo tolerance, faceting, and since 2024 hybrid search with embedders built in, so the keyword and vector halves are one query rather than two systems you reconcile. It is worth reading their own engineering account of building arroy, their LMDB-backed ANN store: it is the clearest public explanation of why hybrid retrieval beats pure vector, and it is a Rust project explaining its own internals.
A lightning-fast search engine API bringing AI-powered hybrid search to your sites and applications.
The dedicated vector database of this shelf, and the one to reach for when the index is the workload rather than a feature of something else. It is one of the few projects here that is Rust all the way down and still keeps a first-class Rust client, which is unusual for a database whose users are mostly writing Python. It is a server you deploy. If you would rather link the index into your own process, LanceDB is the embedded answer; for vectors beside data you already have, sqlite-vec or pgvector.
Qdrant - High-performance, massive-scale Vector Database and Vector Search Engine for the next generation of AI. Also available in the cloud https://cloud.qdrant.io/
An embedded multimodal retrieval database — the one on this shelf you link into your own process rather than deploy next to it, which makes it the natural choice for a Rust service that wants vectors without another server. It sits on the Lance columnar format, so the same data is a table you can scan and a vector index you can search. Rust is a first-class API here, though still 0.x.
LanceDB: A serverless, low-latency vector database for AI applications
Search infrastructure for AI
Ann based on Hierarchical Navigable Small World Graphs from Yu.A. Malkov and D.A Yashunin
pgvector support for Rust
A columnar data format that is 100x faster than Parquet for random access.
Smaller & Faster Single-File Vector Search Engine from Unum
Rust client for Qdrant Vector Search Engine
FFI bindings to the sqlite-vec SQLite extension
Rust wrapper for the Meilisearch API. Meilisearch is a powerful, fast, open-source, easy to use and deploy search engine.
HelixDB is an OLTP graph database with native vector and full-text search built in Rust on Object Storage.
NextPlaid, ColGREP: Multi-vector search, from database to coding agents.
Scalable, fast, and disk-friendly vector search in Postgres, the successor of pgvecto.rs.
Fast vector quantization with 2-4 bit compression and SIMD search
Collected, but not clearing the published listing bar — too new, too quiet, or with no use by anyone but its author that we can see. Plenty of good software is here, some of it years old: this says what we can measure, not what it is worth.
Postgres vector extension, superseded by VectorChord.
Kept so you stop looking for them. Archives are computed on every build and reverse themselves if the reason clears. Where a human has named the reason or a replacement, it cites the public issue behind that.
Archived: repo archived. Detected automatically, and it un-archives itself if that changes.
Fast minimal implementation of HNSW maps for approximate nearest neighbors searches
Turning text into vectors locally, and reordering the results afterwards. Well covered, mostly by running an exported model through one of the inference runtimes.
The default answer to "how do I embed text from Rust". It bundles a model zoo, batching and rerankers over ONNX Runtime, so you get an embedding without assembling a tokenizer, a runtime and a pooling strategy yourself. It is ort underneath, which means a native library alongside your binary; if you want that in a server instead, Hugging Face's text-embeddings-inference is the Rust one.
Library for generating vector embeddings, reranking locally.
A blazing fast inference solution for text embeddings models
Reference model implementations for candle — Llama, Qwen, Gemma, Mistral, Phi, RWKV, BERT/E5/GTE for embeddings, plus Stable Diffusion and Flux — with GGUF and GGML loading.
Portable mixed-precision BLAS-like vector math library for x86 and ARM
Official Rust Implementation of Model2Vec
Highly Performant, Modular, Memory Safe and Production-ready Inference, Ingestion and Indexing built in Rust 🦀
Kept so you stop looking for them. Archives are computed on every build and reverse themselves if the reason clears. Where a human has named the reason or a replacement, it cites the public issue behind that.
Archived: dormant. Detected automatically, and it un-archives itself if that changes.
Reader and writer for common word embedding formats
Keyword search, and the lexical half of hybrid retrieval. Rust-native engines rather than bindings, and mature enough that other search products are built on top of them.
A search engine you can point at a dataset and get relevance from in an afternoon — typo tolerance, faceting, and since 2024 hybrid search with embedders built in, so the keyword and vector halves are one query rather than two systems you reconcile. It is worth reading their own engineering account of building arroy, their LMDB-backed ANN store: it is the clearest public explanation of why hybrid retrieval beats pure vector, and it is a Rust project explaining its own internals.
A lightning-fast search engine API bringing AI-powered hybrid search to your sites and applications.
The Lucene alternative, and the keyword half of hybrid search — which matters because hybrid (BM25 plus vector) is now the default recommendation from the people who build these engines, not pure vector. It is a library, not a server: you get an inverted index, a query parser and fast segment merging, and you decide what to build around them. Quickwit and Meilisearch's lineage come from here.
Search engine library
Collected, but not clearing the published listing bar — too new, too quiet, or with no use by anyone but its author that we can see. Plenty of good software is here, some of it years old: this says what we can measure, not what it is worth.
Chunking, ingesting and keeping a corpus in sync with an index. The chunking primitives are solid; the pipeline frameworks above them are not, so most Rust RAG is assembled by hand out of the shelves either side of this one.
Semantic chunking to a token budget, and the one piece of the RAG pipeline where Rust has a best-in-class answer rather than a catch-up one. It splits on the largest semantic unit that fits — section, paragraph, sentence, word — and takes the length function from tiktoken-rs or tokenizers, so the budget is the model's, not an approximation. Most Rust RAG is assembled by hand out of the shelves either side of this one; this is the part you do not have to write.
Split text into semantic chunks, up to a desired chunk size. Supports calculating length by characters and tokens, and is callable from Rust and Python.
Incremental engine for long horizon agents 🌟 Star if you like it!
A web crawler and scraper, building blocks for data curation workloads.
All-in-one platform for search, recommendations, RAG, and analytics offered via API
Fast, local-first web content extraction for LLMs. Scrape, crawl, extract structured data — all from Rust. CLI, REST API, and MCP server.
Fast, streaming indexing, query, and agentic LLM applications in Rust
Semantic search and document parsing tools for the command line
Turning a PDF, an image or an Office file into text. Extraction and OCR are well covered, including by company-backed Rust parsers; table structure, handwriting and non-Latin scripts are not on this shelf.
A Rust library for PDF document manipulation.
A high-level idiomatic Rust wrapper around Pdfium, the C++ PDF library used by the Google Chromium project.
A library to extract content from pdfs
The fastest Rust PDF library — 0.8ms mean, 5× faster than the industry leaders, 100% pass rate on 3,830 real-world PDFs. Text extraction, Markdown/HTML conversion, PDF creation and editing.
Safe Rust wrapper to MuPDF
OCR engine
Core library for tree-sitter language pack - provides compiled parsers for 371 languages
The fastest Office document processing library — DOCX, XLSX, PPTX, DOC, XLS, PPT
Pure Rust PDF library for AI/RAG: structure-aware chunking, no ML, no C deps.
An Optical Character Recognition (OCR) and Document Layout Analysis library written in Rust.
The v4 LTS line of the project that is now Xberg. Explicitly legacy — new work should use xberg.
Fast, lightweight PDF and document parsing with spatial text extraction
Rust bindings for Tesseract OCR with optional built-in compilation
A lightweight and efficient OCR library based on PaddleOCR models, using the MNN inference framework for high-performance text detection and recognition
Rust multi‑backend OCR/VLM engine (DeepSeek‑OCR-1/2, PaddleOCR‑VL, DotsOCR) with DSQ quantization and an OpenAI‑compatible server & CLI – run locally without Python.
Collected, but not clearing the published listing bar — too new, too quiet, or with no use by anyone but its author that we can see. Plenty of good software is here, some of it years old: this says what we can measure, not what it is worth.
Convert documents (doc, docx, odt, rtf, epub, pdf, presentations, spreadsheets, csv) to GitHub-Flavored Markdown
Higher-level bindings for Tesseract OCR
Modern, fast, document parser written in 🦀
High-performance document intelligence library for Rust. Extract text, metadata, and structured data from PDFs, Office documents, images, and 107 formats and 371 programming languages via tree-sitter code intelligence with async/sync APIs.
Kept so you stop looking for them. Archives are computed on every build and reverse themselves if the reason clears. Where a human has named the reason or a replacement, it cites the public issue behind that.
Are we learning yet?